notesum.ai

Published at November 27

Convolutional Neural Networks Do Work with Pre-Defined Filters

cs.CV

Released Date: November 27, 2024

Authors: Christoph Linse¹, Erhardt Barth¹, Thomas Martinetz¹

Aff.: ¹Institute for Neuro- and Bioinformatics, University of Lübeck, Lübeck, Germany

Arxiv: http://arxiv.org/abs/2411.18388v1

Refer to caption

Layers	Output size	Kernel
Pre-defined Filter Module	$\begin{bmatrix}3\times 224\times 224\\ 48\times 112\times 112\\ 64\times 112\times 112\\ \end{bmatrix}$	$\begin{bmatrix}\text{input},\\ 1\times 3\times 3,\text{depthwise}\\ 48\times 1\times 1,\text{pixelwise}\\ \end{bmatrix}$
MaxPooling	$64\times 56\times 56$	$3\times 3$ , 64, stride 2
Double HF Residual Module	$64\times 56\times 56$	$\begin{bmatrix}1\times 3\times 3,\text{depthwise}\\ 64\times 1\times 1,\text{pixelwise}\\ \end{bmatrix}\times 2$
Double HF Residual Module	$128\times 28\times 28$	$\begin{bmatrix}1\times 3\times 3,\text{depthwise}\\ 64\times 1\times 1,\text{pixelwise}\\ \end{bmatrix}$
Double HF Residual Module	$128\times 28\times 28$	$\begin{bmatrix}1\times 3\times 3,\text{depthwise}\\ 128\times 1\times 1,\text{pixelwise}\\ \end{bmatrix}$
Double HF Residual Module	$256\times 14\times 14$	$\begin{bmatrix}1\times 3\times 3,\text{depthwise}\\ 128\times 1\times 1,\text{pixelwise}\\ \end{bmatrix}$
Double HF Residual Module	$256\times 14\times 14$	$\begin{bmatrix}1\times 3\times 3,\text{depthwise}\\ 256\times 1\times 1,\text{pixelwise}\\ \end{bmatrix}$
Double HF Residual Module	$512\times 7\times 7$	$\begin{bmatrix}1\times 3\times 3,\text{depthwise}\\ 256\times 1\times 1,\text{pixelwise}\\ \end{bmatrix}$
Double HF Residual Module	$512\times 7\times 7$	$\begin{bmatrix}1\times 3\times 3,\text{depthwise}\\ 512\times 1\times 1,\text{pixelwise}\\ \end{bmatrix}$
Classification layer	$512\times 1\times 1$	Adaptive average pool
Classification layer		fully connected, softmax